Merge tag 'ceph-for-6.5-rc1' of https://github.com/ceph/ceph-client
[platform/kernel/linux-starfive.git] / fs / erofs / utils.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2018 HUAWEI, Inc.
4  *             https://www.huawei.com/
5  */
6 #include "internal.h"
7
8 struct page *erofs_allocpage(struct page **pagepool, gfp_t gfp)
9 {
10         struct page *page = *pagepool;
11
12         if (page) {
13                 DBG_BUGON(page_ref_count(page) != 1);
14                 *pagepool = (struct page *)page_private(page);
15         } else {
16                 page = alloc_page(gfp);
17         }
18         return page;
19 }
20
21 void erofs_release_pages(struct page **pagepool)
22 {
23         while (*pagepool) {
24                 struct page *page = *pagepool;
25
26                 *pagepool = (struct page *)page_private(page);
27                 put_page(page);
28         }
29 }
30
31 #ifdef CONFIG_EROFS_FS_ZIP
32 /* global shrink count (for all mounted EROFS instances) */
33 static atomic_long_t erofs_global_shrink_cnt;
34
35 static bool erofs_workgroup_get(struct erofs_workgroup *grp)
36 {
37         if (lockref_get_not_zero(&grp->lockref))
38                 return true;
39
40         spin_lock(&grp->lockref.lock);
41         if (__lockref_is_dead(&grp->lockref)) {
42                 spin_unlock(&grp->lockref.lock);
43                 return false;
44         }
45
46         if (!grp->lockref.count++)
47                 atomic_long_dec(&erofs_global_shrink_cnt);
48         spin_unlock(&grp->lockref.lock);
49         return true;
50 }
51
52 struct erofs_workgroup *erofs_find_workgroup(struct super_block *sb,
53                                              pgoff_t index)
54 {
55         struct erofs_sb_info *sbi = EROFS_SB(sb);
56         struct erofs_workgroup *grp;
57
58 repeat:
59         rcu_read_lock();
60         grp = xa_load(&sbi->managed_pslots, index);
61         if (grp) {
62                 if (!erofs_workgroup_get(grp)) {
63                         /* prefer to relax rcu read side */
64                         rcu_read_unlock();
65                         goto repeat;
66                 }
67
68                 DBG_BUGON(index != grp->index);
69         }
70         rcu_read_unlock();
71         return grp;
72 }
73
74 struct erofs_workgroup *erofs_insert_workgroup(struct super_block *sb,
75                                                struct erofs_workgroup *grp)
76 {
77         struct erofs_sb_info *const sbi = EROFS_SB(sb);
78         struct erofs_workgroup *pre;
79
80         /*
81          * Bump up before making this visible to others for the XArray in order
82          * to avoid potential UAF without serialized by xa_lock.
83          */
84         lockref_get(&grp->lockref);
85
86 repeat:
87         xa_lock(&sbi->managed_pslots);
88         pre = __xa_cmpxchg(&sbi->managed_pslots, grp->index,
89                            NULL, grp, GFP_NOFS);
90         if (pre) {
91                 if (xa_is_err(pre)) {
92                         pre = ERR_PTR(xa_err(pre));
93                 } else if (!erofs_workgroup_get(pre)) {
94                         /* try to legitimize the current in-tree one */
95                         xa_unlock(&sbi->managed_pslots);
96                         cond_resched();
97                         goto repeat;
98                 }
99                 lockref_put_return(&grp->lockref);
100                 grp = pre;
101         }
102         xa_unlock(&sbi->managed_pslots);
103         return grp;
104 }
105
106 static void  __erofs_workgroup_free(struct erofs_workgroup *grp)
107 {
108         atomic_long_dec(&erofs_global_shrink_cnt);
109         erofs_workgroup_free_rcu(grp);
110 }
111
112 void erofs_workgroup_put(struct erofs_workgroup *grp)
113 {
114         if (lockref_put_or_lock(&grp->lockref))
115                 return;
116
117         DBG_BUGON(__lockref_is_dead(&grp->lockref));
118         if (grp->lockref.count == 1)
119                 atomic_long_inc(&erofs_global_shrink_cnt);
120         --grp->lockref.count;
121         spin_unlock(&grp->lockref.lock);
122 }
123
124 static bool erofs_try_to_release_workgroup(struct erofs_sb_info *sbi,
125                                            struct erofs_workgroup *grp)
126 {
127         int free = false;
128
129         spin_lock(&grp->lockref.lock);
130         if (grp->lockref.count)
131                 goto out;
132
133         /*
134          * Note that all cached pages should be detached before deleted from
135          * the XArray. Otherwise some cached pages could be still attached to
136          * the orphan old workgroup when the new one is available in the tree.
137          */
138         if (erofs_try_to_free_all_cached_pages(sbi, grp))
139                 goto out;
140
141         /*
142          * It's impossible to fail after the workgroup is freezed,
143          * however in order to avoid some race conditions, add a
144          * DBG_BUGON to observe this in advance.
145          */
146         DBG_BUGON(__xa_erase(&sbi->managed_pslots, grp->index) != grp);
147
148         lockref_mark_dead(&grp->lockref);
149         free = true;
150 out:
151         spin_unlock(&grp->lockref.lock);
152         if (free)
153                 __erofs_workgroup_free(grp);
154         return free;
155 }
156
157 static unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
158                                               unsigned long nr_shrink)
159 {
160         struct erofs_workgroup *grp;
161         unsigned int freed = 0;
162         unsigned long index;
163
164         xa_lock(&sbi->managed_pslots);
165         xa_for_each(&sbi->managed_pslots, index, grp) {
166                 /* try to shrink each valid workgroup */
167                 if (!erofs_try_to_release_workgroup(sbi, grp))
168                         continue;
169                 xa_unlock(&sbi->managed_pslots);
170
171                 ++freed;
172                 if (!--nr_shrink)
173                         return freed;
174                 xa_lock(&sbi->managed_pslots);
175         }
176         xa_unlock(&sbi->managed_pslots);
177         return freed;
178 }
179
180 /* protected by 'erofs_sb_list_lock' */
181 static unsigned int shrinker_run_no;
182
183 /* protects the mounted 'erofs_sb_list' */
184 static DEFINE_SPINLOCK(erofs_sb_list_lock);
185 static LIST_HEAD(erofs_sb_list);
186
187 void erofs_shrinker_register(struct super_block *sb)
188 {
189         struct erofs_sb_info *sbi = EROFS_SB(sb);
190
191         mutex_init(&sbi->umount_mutex);
192
193         spin_lock(&erofs_sb_list_lock);
194         list_add(&sbi->list, &erofs_sb_list);
195         spin_unlock(&erofs_sb_list_lock);
196 }
197
198 void erofs_shrinker_unregister(struct super_block *sb)
199 {
200         struct erofs_sb_info *const sbi = EROFS_SB(sb);
201
202         mutex_lock(&sbi->umount_mutex);
203         /* clean up all remaining workgroups in memory */
204         erofs_shrink_workstation(sbi, ~0UL);
205
206         spin_lock(&erofs_sb_list_lock);
207         list_del(&sbi->list);
208         spin_unlock(&erofs_sb_list_lock);
209         mutex_unlock(&sbi->umount_mutex);
210 }
211
212 static unsigned long erofs_shrink_count(struct shrinker *shrink,
213                                         struct shrink_control *sc)
214 {
215         return atomic_long_read(&erofs_global_shrink_cnt);
216 }
217
218 static unsigned long erofs_shrink_scan(struct shrinker *shrink,
219                                        struct shrink_control *sc)
220 {
221         struct erofs_sb_info *sbi;
222         struct list_head *p;
223
224         unsigned long nr = sc->nr_to_scan;
225         unsigned int run_no;
226         unsigned long freed = 0;
227
228         spin_lock(&erofs_sb_list_lock);
229         do {
230                 run_no = ++shrinker_run_no;
231         } while (run_no == 0);
232
233         /* Iterate over all mounted superblocks and try to shrink them */
234         p = erofs_sb_list.next;
235         while (p != &erofs_sb_list) {
236                 sbi = list_entry(p, struct erofs_sb_info, list);
237
238                 /*
239                  * We move the ones we do to the end of the list, so we stop
240                  * when we see one we have already done.
241                  */
242                 if (sbi->shrinker_run_no == run_no)
243                         break;
244
245                 if (!mutex_trylock(&sbi->umount_mutex)) {
246                         p = p->next;
247                         continue;
248                 }
249
250                 spin_unlock(&erofs_sb_list_lock);
251                 sbi->shrinker_run_no = run_no;
252
253                 freed += erofs_shrink_workstation(sbi, nr - freed);
254
255                 spin_lock(&erofs_sb_list_lock);
256                 /* Get the next list element before we move this one */
257                 p = p->next;
258
259                 /*
260                  * Move this one to the end of the list to provide some
261                  * fairness.
262                  */
263                 list_move_tail(&sbi->list, &erofs_sb_list);
264                 mutex_unlock(&sbi->umount_mutex);
265
266                 if (freed >= nr)
267                         break;
268         }
269         spin_unlock(&erofs_sb_list_lock);
270         return freed;
271 }
272
273 static struct shrinker erofs_shrinker_info = {
274         .scan_objects = erofs_shrink_scan,
275         .count_objects = erofs_shrink_count,
276         .seeks = DEFAULT_SEEKS,
277 };
278
279 int __init erofs_init_shrinker(void)
280 {
281         return register_shrinker(&erofs_shrinker_info, "erofs-shrinker");
282 }
283
284 void erofs_exit_shrinker(void)
285 {
286         unregister_shrinker(&erofs_shrinker_info);
287 }
288 #endif  /* !CONFIG_EROFS_FS_ZIP */